huihui-ai/Huihui-gemma-4-31B-it-abliterated-v2
The huihui-ai/Huihui-gemma-4-31B-it-abliterated-v2 is a 31 billion parameter instruction-tuned causal language model, derived from Google's Gemma-4-31B-it. This model has been modified using 'abliteration' techniques to significantly reduce safety filtering and refusals, resulting in a lower perplexity (PPL) value compared to its base model. It is optimized for generating less censored and more direct outputs, making it suitable for research and experimental use cases where content filtering is undesirable.
Loading preview...
Model Overview
The huihui-ai/Huihui-gemma-4-31B-it-abliterated-v2 is a 31 billion parameter instruction-tuned model based on google/gemma-4-31B-it. Its primary distinction lies in the application of "abliteration" techniques, specifically designed to remove refusal behaviors and reduce safety filtering present in the original model. This modification aims to provide a more direct and less censored output experience.
Key Characteristics
- Reduced Refusals: Modified to minimize safety filtering and refusal responses, offering uncensored outputs.
- Improved Perplexity (PPL): Demonstrates a lower perplexity value (13161.2940) compared to the base
google/gemma-4-31B-it(14874.7532), indicating potentially better language modeling quality. - Context Length: Supports a context length of 32768 tokens.
- Proof-of-Concept: Represents a proof-of-concept implementation for removing refusals without relying on TransformerLens.
Usage Considerations
This model is explicitly noted for its significantly reduced safety filtering, which means it may generate sensitive, controversial, or inappropriate content. Users are advised to:
- Exercise caution and rigorously review all generated outputs.
- Avoid use in public-facing commercial applications or for underage users.
- Ensure compliance with local laws and ethical standards.
- Primarily use for research, testing, or controlled environments.